During a call via a communication device, contamination of environmental noise into speech collected at a microphone of the device reduces the rate of successful speech recognition. In order to prevent such a reduction in the rate of successful speech recognition, in a transmission mode, the noise contained in a speech signal must be attenuated before transmission of the speech signal. In addition, in a reception mode, the noise contained in a reception signal must be attenuated before output of the reception signal.
A speech recognition device disclosed in Unexamined Japanese Patent Application Kokai Publication No. 1999-231897 analyzes an input signal from a single microphone and determines the spectrum of the signal immediately before a speech period to be a noise spectrum. The speech recognition device subtracts the noise spectrum from the spectrum within the entire speech period to acquire the spectrum within the speech period without noise. A speech processing apparatus disclosed in Unexamined Japanese Patent Application Kokai Publication No. 2012-133346 calculates the total energy of a spectrum group containing a given spectrum and neighboring spectra in a spectral pattern, and sequentially calculates the total energies while shifting the focus on the spectrum one by one. The speech processing apparatus then detects a peak spectrum based on the energy ratio between the total energy of a target spectrum group and the total energy of an adjacent spectrum group. The speech processing apparatus uses the barycentric frequency of the peak spectrum to determine the noise spectrum depending on the number of barycentric frequencies having a harmonic relationship, and thus suppresses the noise spectrum.